I was in more meetings than usual today so I just caught up with the news that Cohere, the Canadian startup co-founded by former Transformer paper author Aidan Gomez, has released its first reasoning large language model (LLM), Command A Reasoning. This model aims to make generative AI products work easily, powerfully, and securely for enterprises.
This is a strong release, according to benchmarks, technical specs, and early tests that suggest flexibility, efficiency, and raw reasoning power. Cohere claims it can handle customer service, market research, scheduling, and data analysis tasks automatically at scale within secure enterprise environments.
Although it is a text-only model, it can connect to multimodal models and tools. Tool use is one of its primary selling points. The model is open for researchers for non-commercial use, but enterprises will need to pay for access as Cohere offers bespoke customization and private deployment.
Cohere was valued at $6.8 billion when it announced its latest funding round of $500 million recently.
Command A Reasoning is designed for enterprises with extensive document libraries, long email chains, and workflows that cannot afford hallucinations. It supports 256,000 tokens on multi-GPU setups, comparable to OpenAI’s GPT-5. The research release has 111-billion parameters, trained with tool-use and multilingual performance, supporting 23 languages like English, French, Spanish, Japanese, Arabic, and Hindi.
The model integrates into North, Cohere’s platform for AI agents and automations on-premises. Enterprises can create custom agents within their infrastructure, maintaining data control while leveraging advanced reasoning.
Cohere seems to have cleverly designed its model to support recurring enterprise functions such as onboarding, market research, and development. Command A Reasoning introduces a token budget feature for users or developers to allocate specific reasoning to tasks. Developers can operate the model in “reasoning mode” for high performance or turn it off for less intensive tasks.
Cohere’s benchmarks show that Command A Reasoning consistently outperforms competitors like DeepSeek-R1 0528, gpt-oss-120b, and Mistral Magistral Medium in enterprise reasoning tasks and multilingual benchmarks. Satisfaction scores increase with token budget, even with minimal reasoning, it surpasses its predecessor.
Aside from benchmarks, the model is trained for conversational tool use, allowing it to call APIs, connect to databases, or query systems during tasks. Developers can define tools via JSON schema in Transformers, facilitating integration into enterprise systems.
Safety is also a highlight, with training to avoid over-refusal while filtering harmful content. Evaluations focus on five high-risk categories: child safety, self-harm, violence and hate, explicit material, and conspiracy theories.
SAP SE is an early adopter, integrating the model to enhance generative AI capabilities within its SAP Business Technology Platform, allowing for customizable agentic applications.
Command A Reasoning is now available on the Cohere platform and for research on Hugging Face. The Hugging Face repository offers open weights for research under a CC-BY-NC license. Enterprises interested in commercial or private deployments can contact Cohere for bespoke pricing.
In summary, Command A Reasoning offers one model with various deployment modes, fine control over performance, multilingual capability, tool integration, and benchmark results suggesting superiority over peers.